Examples with SparkRDDWriteClient org.apache.hudi.client.SparkRDDWriteClient used on opensource projects

Example 51 with SparkRDDWriteClient

use of org.apache.hudi.client.SparkRDDWriteClient in project hudi by apache.

the class TestHoodieBackedMetadata method testMetadataMultiWriter.

/**
 * Test multi-writer on metadata table with optimistic concurrency.
 */
@Test
public void testMetadataMultiWriter() throws Exception {
    init(HoodieTableType.COPY_ON_WRITE);
    HoodieSparkEngineContext engineContext = new HoodieSparkEngineContext(jsc);
    Properties properties = new Properties();
    properties.setProperty(FILESYSTEM_LOCK_PATH_PROP_KEY, basePath + "/.hoodie/.locks");
    properties.setProperty(LockConfiguration.LOCK_ACQUIRE_WAIT_TIMEOUT_MS_PROP_KEY, "1000");
    properties.setProperty(LockConfiguration.LOCK_ACQUIRE_CLIENT_NUM_RETRIES_PROP_KEY, "20");
    HoodieWriteConfig writeConfig = getWriteConfigBuilder(true, true, false).withCompactionConfig(HoodieCompactionConfig.newBuilder().withFailedWritesCleaningPolicy(HoodieFailedWritesCleaningPolicy.LAZY).withAutoClean(false).build()).withWriteConcurrencyMode(WriteConcurrencyMode.OPTIMISTIC_CONCURRENCY_CONTROL).withLockConfig(HoodieLockConfig.newBuilder().withLockProvider(InProcessLockProvider.class).build()).withProperties(properties).build();
    ExecutorService executors = Executors.newFixedThreadPool(dataGen.getPartitionPaths().length);
    // Create clients in advance
    SparkRDDWriteClient[] writeClients = new SparkRDDWriteClient[dataGen.getPartitionPaths().length];
    for (int i = 0; i < dataGen.getPartitionPaths().length; i++) {
        writeClients[i] = new SparkRDDWriteClient(engineContext, writeConfig);
    }
    // Parallel commits for separate partitions
    List<Future> futures = new LinkedList<>();
    for (int i = 0; i < dataGen.getPartitionPaths().length; ++i) {
        final int index = i;
        String newCommitTime = "000000" + (index + 1);
        Future future = executors.submit(() -> {
            List<HoodieRecord> records = dataGen.generateInsertsForPartition(newCommitTime, 100, dataGen.getPartitionPaths()[index]);
            SparkRDDWriteClient writeClient = writeClients[index];
            writeClient.startCommitWithTime(newCommitTime);
            List<WriteStatus> writeStatuses = writeClient.insert(jsc.parallelize(records, 1), newCommitTime).collect();
            assertNoWriteErrors(writeStatuses);
        });
        futures.add(future);
    }
    // Wait for all commits to complete
    for (Future future : futures) {
        future.get();
    }
    // Ensure all commits were synced to the Metadata Table
    HoodieTableMetaClient metadataMetaClient = HoodieTableMetaClient.builder().setConf(hadoopConf).setBasePath(metadataTableBasePath).build();
    assertEquals(metadataMetaClient.getActiveTimeline().getDeltaCommitTimeline().filterCompletedInstants().countInstants(), 4);
    assertTrue(metadataMetaClient.getActiveTimeline().containsInstant(new HoodieInstant(false, HoodieTimeline.DELTA_COMMIT_ACTION, "0000001")));
    assertTrue(metadataMetaClient.getActiveTimeline().containsInstant(new HoodieInstant(false, HoodieTimeline.DELTA_COMMIT_ACTION, "0000002")));
    assertTrue(metadataMetaClient.getActiveTimeline().containsInstant(new HoodieInstant(false, HoodieTimeline.DELTA_COMMIT_ACTION, "0000003")));
    // Compaction may occur if the commits completed in order
    assertTrue(metadataMetaClient.getActiveTimeline().getCommitTimeline().filterCompletedInstants().countInstants() <= 1);
    // Validation
    validateMetadata(writeClients[0]);
}

Example 52 with SparkRDDWriteClient

use of org.apache.hudi.client.SparkRDDWriteClient in project hudi by apache.

the class TestHoodieBackedMetadata method testReattemptOfFailedClusteringCommit.

/**
 * Lets say clustering commit succeeded in metadata table, but failed before committing to datatable.
 * Next time, when clustering kicks in, hudi will rollback pending clustering (in data table) and re-attempt the clustering with same
 * instant time. So, this test ensures the 2nd attempt succeeds with metadata enabled.
 * This is applicable to any table service where instant time is fixed. So, how many ever times the operation fails, re attempt will
 * be made with same commit time.
 * Tests uses clustering to test out the scenario.
 */
@Test
public void testReattemptOfFailedClusteringCommit() throws Exception {
    tableType = HoodieTableType.COPY_ON_WRITE;
    init(tableType);
    context = new HoodieSparkEngineContext(jsc);
    HoodieWriteConfig config = getSmallInsertWriteConfig(2000, TRIP_EXAMPLE_SCHEMA, 10, false);
    SparkRDDWriteClient client = getHoodieWriteClient(config);
    // Write 1 (Bulk insert)
    String newCommitTime = "0000001";
    List<HoodieRecord> records = dataGen.generateInserts(newCommitTime, 20);
    client.startCommitWithTime(newCommitTime);
    List<WriteStatus> writeStatuses = client.insert(jsc.parallelize(records, 1), newCommitTime).collect();
    assertNoWriteErrors(writeStatuses);
    validateMetadata(client);
    // Write 2 (inserts)
    newCommitTime = "0000002";
    client.startCommitWithTime(newCommitTime);
    records = dataGen.generateInserts(newCommitTime, 20);
    writeStatuses = client.insert(jsc.parallelize(records, 1), newCommitTime).collect();
    assertNoWriteErrors(writeStatuses);
    validateMetadata(client);
    // setup clustering config.
    HoodieClusteringConfig clusteringConfig = HoodieClusteringConfig.newBuilder().withClusteringMaxNumGroups(10).withClusteringSortColumns("_row_key").withInlineClustering(true).withClusteringTargetPartitions(0).withInlineClusteringNumCommits(1).build();
    HoodieWriteConfig newWriteConfig = getConfigBuilder(TRIP_EXAMPLE_SCHEMA, HoodieIndex.IndexType.BLOOM, HoodieFailedWritesCleaningPolicy.EAGER).withAutoCommit(false).withClusteringConfig(clusteringConfig).build();
    // trigger clustering
    SparkRDDWriteClient newClient = getHoodieWriteClient(newWriteConfig);
    String clusteringCommitTime = newClient.scheduleClustering(Option.empty()).get().toString();
    HoodieWriteMetadata<JavaRDD<WriteStatus>> clusterMetadata = newClient.cluster(clusteringCommitTime, true);
    // collect replaceFileIds for validation later.
    Set<HoodieFileGroupId> replacedFileIds = new HashSet<>();
    clusterMetadata.getPartitionToReplaceFileIds().entrySet().forEach(partitionFiles -> partitionFiles.getValue().stream().forEach(file -> replacedFileIds.add(new HoodieFileGroupId(partitionFiles.getKey(), file))));
    // trigger new write to mimic other writes succeeding before re-attempt.
    newCommitTime = "0000003";
    client.startCommitWithTime(newCommitTime);
    records = dataGen.generateInserts(newCommitTime, 20);
    writeStatuses = client.insert(jsc.parallelize(records, 1), newCommitTime).collect();
    assertNoWriteErrors(writeStatuses);
    validateMetadata(client);
    // manually remove clustering completed instant from .hoodie folder and to mimic succeeded clustering in metadata table, but failed in data table.
    FileCreateUtils.deleteReplaceCommit(basePath, clusteringCommitTime);
    HoodieWriteMetadata<JavaRDD<WriteStatus>> updatedClusterMetadata = newClient.cluster(clusteringCommitTime, true);
    metaClient.reloadActiveTimeline();
    Set<HoodieFileGroupId> updatedReplacedFileIds = new HashSet<>();
    updatedClusterMetadata.getPartitionToReplaceFileIds().entrySet().forEach(partitionFiles -> partitionFiles.getValue().stream().forEach(file -> updatedReplacedFileIds.add(new HoodieFileGroupId(partitionFiles.getKey(), file))));
    assertEquals(replacedFileIds, updatedReplacedFileIds);
    validateMetadata(client);
}

Example 53 with SparkRDDWriteClient

use of org.apache.hudi.client.SparkRDDWriteClient in project hudi by apache.

the class TestHoodieBackedMetadata method testRollbackOfPartiallyFailedCommitWithNewPartitions.

/**
 * Tests rollback of a commit which has new partitions which is not present in hudi table prior to the commit being rolledback.
 *
 * @throws Exception
 */
@Test
public void testRollbackOfPartiallyFailedCommitWithNewPartitions() throws Exception {
    init(HoodieTableType.COPY_ON_WRITE);
    HoodieSparkEngineContext engineContext = new HoodieSparkEngineContext(jsc);
    try (SparkRDDWriteClient client = new SparkRDDWriteClient(engineContext, getWriteConfigBuilder(HoodieFailedWritesCleaningPolicy.EAGER, true, true, false, true, false, false).build(), true)) {
        String newCommitTime = HoodieActiveTimeline.createNewInstantTime();
        client.startCommitWithTime(newCommitTime);
        List<HoodieRecord> records = dataGen.generateInserts(newCommitTime, 10);
        List<HoodieRecord> upsertRecords = new ArrayList<>();
        for (HoodieRecord entry : records) {
            if (entry.getPartitionPath().equals(HoodieTestDataGenerator.DEFAULT_FIRST_PARTITION_PATH) || entry.getPartitionPath().equals(HoodieTestDataGenerator.DEFAULT_SECOND_PARTITION_PATH)) {
                upsertRecords.add(entry);
            }
        }
        List<WriteStatus> writeStatuses = client.upsert(jsc.parallelize(upsertRecords, 1), newCommitTime).collect();
        assertNoWriteErrors(writeStatuses);
        validateMetadata(client);
        newCommitTime = HoodieActiveTimeline.createNewInstantTime();
        client.startCommitWithTime(newCommitTime);
        records = dataGen.generateInserts(newCommitTime, 20);
        writeStatuses = client.insert(jsc.parallelize(records, 1), newCommitTime).collect();
        assertNoWriteErrors(writeStatuses);
        validateMetadata(client);
        // There is no way to simulate failed commit on the main dataset, hence we simply delete the completed
        // instant so that only the inflight is left over.
        String commitInstantFileName = HoodieTimeline.makeCommitFileName(newCommitTime);
        assertTrue(fs.delete(new Path(basePath + Path.SEPARATOR + HoodieTableMetaClient.METAFOLDER_NAME, commitInstantFileName), false));
    }
    try (SparkRDDWriteClient client = new SparkRDDWriteClient(engineContext, getWriteConfigBuilder(HoodieFailedWritesCleaningPolicy.EAGER, true, true, false, true, false, false).build(), true)) {
        String newCommitTime = client.startCommit();
        // Next insert
        List<HoodieRecord> records = dataGen.generateInserts(newCommitTime, 20);
        List<WriteStatus> writeStatuses = client.upsert(jsc.parallelize(records, 1), newCommitTime).collect();
        assertNoWriteErrors(writeStatuses);
        validateMetadata(client);
    }
}

Example 54 with SparkRDDWriteClient

use of org.apache.hudi.client.SparkRDDWriteClient in project hudi by apache.

the class TestHoodieBackedMetadata method testFirstCommitRollback.

// Some operations are not feasible with test table infra. hence using write client to test those cases.
/**
 * Rollback of the first commit should not trigger bootstrap errors at the metadata table.
 */
@ParameterizedTest
@EnumSource(HoodieTableType.class)
public void testFirstCommitRollback(HoodieTableType tableType) throws Exception {
    init(tableType);
    HoodieSparkEngineContext engineContext = new HoodieSparkEngineContext(jsc);
    try (SparkRDDWriteClient client = new SparkRDDWriteClient(engineContext, getWriteConfigBuilder(true, true, false).withRollbackUsingMarkers(false).build())) {
        // Write 1
        String commitTime = "0000001";
        List<HoodieRecord> records = dataGen.generateInserts(commitTime, 20);
        client.startCommitWithTime(commitTime);
        List<WriteStatus> writeStatuses = client.insert(jsc.parallelize(records, 1), commitTime).collect();
        assertNoWriteErrors(writeStatuses);
        validateMetadata(client);
        // Rollback the first commit
        client.rollback(commitTime);
        // Write 2
        commitTime = "0000002";
        records = dataGen.generateInserts(commitTime, 10);
        client.startCommitWithTime(commitTime);
        writeStatuses = client.upsert(jsc.parallelize(records, 1), commitTime).collect();
        assertNoWriteErrors(writeStatuses);
        validateMetadata(client);
    }
}

Example 55 with SparkRDDWriteClient

use of org.apache.hudi.client.SparkRDDWriteClient in project hudi by apache.

the class TestHoodieClientOnCopyOnWriteStorage method testDeletesWithoutInserts.

/**
 * Test delete with delete api.
 */
@ParameterizedTest
@MethodSource("populateMetaFieldsParams")
public void testDeletesWithoutInserts(boolean populateMetaFields) {
    final String testPartitionPath = "2016/09/26";
    final int insertSplitLimit = 100;
    // setup the small file handling params
    HoodieWriteConfig config = getSmallInsertWriteConfig(insertSplitLimit, TRIP_EXAMPLE_SCHEMA, dataGen.getEstimatedFileSizeInBytes(150), populateMetaFields, populateMetaFields ? new Properties() : getPropertiesForKeyGen());
    dataGen = new HoodieTestDataGenerator(new String[] { testPartitionPath });
    SparkRDDWriteClient client = getHoodieWriteClient(config);
    // delete non existent keys
    String commitTime1 = "001";
    client.startCommitWithTime(commitTime1);
    List<HoodieRecord> dummyInserts = dataGen.generateInserts(commitTime1, 20);
    List<HoodieKey> hoodieKeysToDelete = randomSelectAsHoodieKeys(dummyInserts, 20);
    JavaRDD<HoodieKey> deleteKeys = jsc.parallelize(hoodieKeysToDelete, 1);
    assertThrows(HoodieIOException.class, () -> {
        client.delete(deleteKeys, commitTime1).collect();
    }, "Should have thrown Exception");
}